Research Of Multi-frame Low-light Imaging With An Improved U-net

Photos have become an important medium for people to record their lives and share with each other.As clear images can better express the information conveyed by the photographer,people put forward higher requirements for photo quality.However,in the low-light environment,photos often have problems such as loss of details and serious noise,which will seriously affect the visual feelings of the image.This thesis propose a multi-frame image fusion model named MSID to enhance the visibility and contrast of low-light images,which has very important research significance and practical value in the field of imaging.In order to optimize the imaging effect of low-light images,this thesis firstly analyzes relevant algorithms of low-light image processing and focuses on the SID single-frame image enhancement algorithm based on Raw data.Then,aiming at the problems of spots and color blocks in the image processing by SID algorithm,a multi-frame fusion network structure with HDC module is proposed,which can efficiently aggregate the global information of an image.And the combined loss function is used to train the model,which can effectively improve the model’s performance.At the same time,for the problem that SID algorithm takes up a large amount of memory and costs a long time to calculate,the depthwise convolution and inverted residuals in Mobile Net networks are used to effectively reduce the model parameters and shorten the computation time.Finally,an end-to-end model MSID for low-light imaging from Raw image data to the RGB images is obtained.Then the algorithm is deployed on the cell phone through Tensorflow Lite,and the practicability of the MSID model is further verified by the actual photo test.Based on the SID dataset,this thesis compares the proposed MSID model with the SID model and the Fast DVDNet under the same conditions.Compared with the SID model,the PSNR and SSIM of the dual-frame image fusion MSID model are improved to 28.786 d B and0.887.At the same time,parameters of the model are reduced by 93% and FLOPs is reduced by 87%.The PSNR and SSIM of the five-frame fusion imaging MSID model are 29.668 d B and 0.902,respectively,which are higher than the results of the Fast DVDNet,while parameters and FLOPs only accounted for 22% and 6% of the Fast DVDNet.Experimental results show that the proposed MSID algorithm has better advantages in imaging quality and computing time.